Benchmarking Uncertainty Calibration in Large Language Model Long-Form Question Answering

TL;DR

This paper introduces a comprehensive benchmark to evaluate uncertainty quantification methods in large language models for scientific question answering, revealing significant limitations in current approaches and measurement practices.

Contribution

It provides the first large-scale, open-source framework for assessing UQ calibration in reasoning-intensive scientific QA across multiple models and datasets.

Findings

Instruction tuning causes probability mass polarization, reducing token-level confidence reliability.

Verbalized UQ approaches are biased and poorly correlated with correctness.

Answer frequency is the most reliable calibration method among those tested.

Abstract

Large Language Models (LLMs) are commonly used in Question Answering (QA) settings, increasingly in the natural sciences if not science at large. Reliable Uncertainty Quantification (UQ) is critical for the trustworthy uptake of generated answers. Existing UQ approaches remain weakly validated in scientific QA, a domain relying on fact-retrieval and reasoning capabilities. We introduce the first large-scale benchmark for evaluating UQ metrics in reasoning-demanding QA studying calibration of UQ methods, providing an extensible open-source framework to reproducibly assess calibration. Our study spans up to 20 large language models of base, instruction-tuned and reasoning variants. Our analysis covers seven scientific QA datasets, including both multiple-choice and arithmetic question answering tasks, using prompting to emulate an open question answering setting. We evaluate and compare methods representative of prominent approaches on a total of 685,000 long-form responses, spanning different reasoning complexities representative of domain-specific tasks. At the token level, we find that instruction tuning induces strong probability mass polarization, reducing the reliability of token-level confidences as estimates of uncertainty. Models further fine-tuned for reasoning are exposed to the same effect, but the reasoning process appears to mitigate it depending on the provider. At the sequence level, we show that verbalized approaches are systematically biased and poorly correlated with correctness, while answer frequency (consistency across samples) yields the most reliable calibration. In the wake of our analysis, we study and report the misleading effect of relying exclusively on ECE as a sole measure for judging performance of UQ methods on benchmark datasets. Our findings expose critical limitations of current UQ methods for LLMs and standard practices in benchmarking thereof.